[computer-go] Optimal explore rates for plain UCT

[Petr Baudis]

On Thu, Mar 13, 2008 at 09:19:58AM -0400, Don Dailey wrote:
> 
> > If it is agreed,  I will start a 25k test.    My prediction is that this
> > will finish around 1600 ELO on CGOS.   
> >   
> Looks like it's currently around 1485,  so I am 115 ELO off from my
> prediction at the moment.
> 
> 2 more doublings would probably get you to 1700 -  perhaps I will test
> this later,  I think I can handle 100,000 nodes without a problem timing
> out.      If it's too much I think I can solve it by playing quickly
> when the score is extreme - and it will have little impact on the strength.

I stop playouts when the best candidate has probability >= 0.95 and has
seen at least 500 playouts; this seems also much friendlier to humans
when playing on KGS. Watching many games, I didn't see any fluke about
this when raising the bar to 500 playouts (it was 100 playouts before
and I have seen one or two), then again my bot has lower ELO on CGOS
than it should for its number of playouts, so... ;-)

-- 
				Petr "Pasky" Baudis
Whatever you can do, or dream you can, begin it.
Boldness has genius, power, and magic in it.	-- J. W. von Goethe